In view of the long training time, easy overfitting and sensitive to noise of the transformer fault diagnosis model, a deep denoising extreme learning machine method of transformer fault diagnosis is proposed in this paper. Combine extreme learning machine and denoising auto-encoder to build denoising auto-encoding extreme learning machine, which is stacked to extract feature and an original extreme learning machine is to classify. Both of them form deep denoising extreme learning machine classification algorithm. The algorithm can effectively deal with noise in DGA data and has a very fast learning speed. The simulation experiment result shows that compared with the BP method, the method in this paper has higher fault diagnosis accuracy and shorter training time. It is an effective method of transformer fault diagnosis.